Unpaired Bone Marrow Smears Virtual Staining Using Content and Attention-Guided Generative Adversarial Networks UBMSVStain Using Content and Attention-Guided Generative Adversarial Networks

Morphological examination of bone marrow cells is the standard for the diagnosis of Acute Lymphoblastic Leukemia (ALL). However, the preparation of bone marrow smears requires tedious staining steps, trained pathologists, and a professional experimental environment. Hence, a method that can transform bright-field microscopy images of unstained bone marrow cell smears into Wright & Giemsa (W&G)-stained images of the same samples is essential. However, paired image data is hardly available. In addition, existing unsupervised methods have limitations on datasets of entirely unpaired and complex texture images. This paper proposes an Unpaired Bone Marrow Smears Virtual Staining (UBMSVStain) method, in which a Content And Attention-Guided Staining (CAGS) module is designed to enhance the features after the skip connections and improve the preservation of structural information. All experimental results show that UBMSVStain not only achieves virtual staining of bone marrow smears but also has superior performance.